Physics-based machine learning for enhanced drug formulation development | JoVE Visualize

Area of Science:

Pharmaceutical Sciences
Computational Chemistry
Artificial Intelligence in Drug Development

Background:

Drug formulation development faces challenges due to limited and varied experimental data, hindering conventional AI model accuracy and generalizability.
Accurate prediction of formulation properties like physical stability and hygroscopicity is crucial for successful drug development.

Purpose of the Study:

To introduce and validate a physics-based machine learning (PBML) framework that integrates physics-based modeling with data-driven learning for enhanced drug formulation design.
To improve the prediction of physical stability for amorphous solid dispersions (ASDs) and molecular hygroscopicity using PBML.

Main Methods:

Developed a PBML approach combining molecular dynamics (MD) simulations with machine learning models (TabPFN).
Extracted MD-derived descriptors (e.g., drug-polymer interactions, diffusion coefficients, surface polarity, electrostatic potential variance) to represent formulation properties.
Validated the model's performance on predicting ASD physical stability and molecular hygroscopicity using experimental datasets and SHapley Additive exPlanations (SHAP) for interpretability.

Main Results:

MD-derived descriptors significantly outperformed empirical parameters for ASD stability prediction (75.2% vs. 66.1% generalization).
The hygroscopicity classification model achieved high accuracy (0.967), F1-score (0.957), and AUC-ROC (0.931) on external experimental data.
SHAP analyses confirmed that key features like drug-polymer attraction, API mobility, surface polarity, and ESP variance align with known experimental mechanisms, enhancing model interpretability.

Conclusions:

The PBML framework demonstrates a data-efficient and mechanism-grounded approach for drug formulation design, effectively utilizing limited datasets.
This approach enhances decision-making in formulation development, improving the prediction of critical properties and reducing experimental burden.
The integration of physics-based insights with machine learning offers a powerful strategy for advancing pharmaceutical research and development.